IEEE Robotics & Automation Magazine - June 2011 - 88

Symbolic Processing
Event-Driven Processing

High-Level
Activity Models

t

A6,P6,t6
A4,P4,t4

A5,P5,t5
A2,P2,t2

A1,P1,t1

A3,P3,t3

Reasoning Engine

{(d1,p1),...}

Decision Filtering
Fusion

{(c2,1,p2,1),...}

{(c3,1,p3,1),...}
C3

Classification

F

F

F

Feature Extraction

S

S

S

S

P

P

P

P

Segmentation

P

P

S

Wi

F

Xi

C2

C1

{(c1,1,p1,1),...}

Decision Fusion

Preprocessing

S
Sensor Acquisition

Body-Worn
Sensors

Raw Sensor Data
IEEE ROBOTICS & AUTOMATION MAGAZINE

*

JUNE 2011

Subsymbolic Processing
Streaming Signal Processing

Low-Level Activity Models
(Action Primitives)

{(Ai,Pi,ti),...}
Null Class Rejection

Activity-Recognition Chain

Action Primitives

*

Figure 5. Processing steps used to infer activities from on-body sensors. The raw sensor data are mapped to the occurrence of action primitive (events) with signal processing and
machine-learning techniques. Here, five sensors deliver data. Data fusion is illustrated at the feature, classifier, and decision levels. Symbolic processing infers higher-level activities
from the occurrence of action primitives, usually with reasoning or statistical approaches.

Activity-Aware
Application

Context, Activity, Behavior
88

activities or reject the null class
or a set of features, are also optimized before operation.
Classifiers commonly used for
activity recognition have been
reviewed in [38] together with the
typical features derived from
acceleration signals. If the features corresponding to activities
form clusters in the feature space
(see Figure 4), then the classifiers
that are typically used include
support vector machines [40],
decision trees, k-nearest neighbor,
or Naive Bayes classifiers [41].
This is usually the case with isolated gestures and when static
postures are recognized with features such as limb angles. It is also
the case with periodic activities
when frequency-domain features
are used (e.g., walking leads to
energy in specific frequency bands).
When the temporal unfolding of
the gesture is analyzed, such as
sporadic gestures, approaches such
as dynamic time warping [42] or
hidden Markov models (HMMs)
[25] are used. Other methods
include neural networks [43] or
fuzzy systems [44].
In Figure 4, we illustrate a set
of activities and its corresponding
sensor signals. We can note the
variability in the gesture execution
length and signal shape. With
simple statistical features, the sensor signals can be projected in a
feature space where the activities
form clusters suitable for classification. During the training of the
recognition chain, the selection of
preprocessing steps and features
aims at increasing the separation
between the activity classes. Some
activities are well separated, leading to accurate classification (e.g.,
C10, C1), while others overlap as
they are more similar (e.g., C5,
C6, bottom right).
Symbolic-level processing is
usually event driven, with the
events corresponding to activity
occurrences. Higher-level activity models are thus built on event
occurrences instead of raw
Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2011
